Slashdot Mirror
SGI Demos 64-Proc Linux Box
foobar104 writes "Details are scarce, but SGI announced this morning that their prototype Itanium 2 system has demonstrated more than 120 GB/s to and from main memory on the STREAM TRIAD benchmark, which is the fourth best result in the world. For comparison, the Cray C90 sustains 105 GB/s, while an even larger Sun Fire 15K clocks a measly 55 GB/s. The interesting part? The system wasn't running IRIX, SGI's proprietary version of UNIX. It was running Linux. More information on STREAM TRIAD, including results from other systems, is available here. The system, incidentally, was an Origin 3800 straight out of manufacturing equipped with Itanium 2 processor modules. SGI will start selling the systems early next year."
That was my first though. So it beats a C90, but what is faster?
Found the answer here.
And if you were wondering about a Beowolf cluster of these, the top ten ranking excludes "cluster results".
Linux running at 120 GB/s with 64 processors is impressive for an OS that has been criticized as inefficient when running on more than 8.
I would be very interested to know what version of the kernel they are using.
Typical super-computing problems are weather prediction, air flow computations and nuclear reaction modelling. Physical models in other words.
Generally, you attack these kinds of problem by partitioning 3-d space into many small cells, and then running relatively simple calculations on every cell. The better the resolution, the better the model.
The thing about three dimensions is, storage space increases with resolution^3... For instance, I believe the weather guys are currently pushing 1kmx1kmx100m resolutions. That means about 3,2e11 cells. If each cell has 1 kB of state, the total memory usage would be about 320 TB.
Super computing problems eat memory like Takeru Kobayashi eats hot dogs. In many (most?) cases the calculations are simple. Hence, bandwidth is King.
I choose to remain celibate, like my father and his father before him.
One of the areas this is meaningful is data warehousing. There are three major competitors in the very large data warehousing environment and one wanna be competitor:
- NCR Teradata and Worldmark MPP servers
- IBM DB2 and IBM pSeries clusters (MPP again)
- Sun SunFire 15K and Sybase IQ Multiplex (SMP)
- Oracle is trying to compete in this space and not really succeeding. Their model is sort of MPP, based on Oracle Real Application Clusters
MPP, or massively parallel processing, is the typical solution for very large (generally anything over 3 or 4 terabytes) data warehouses. Sun and Sybase are trying hard to crack the market with their SMP (symmetric multi-processing) solution, which is actually very promising. The major benefit to SMP processing is simplicity, one server to maintain, one OS, no cluster, no cluster interconnect. With Linux potentially pushing into the large SMP space we will have the potential for competition to the MPP data warehouse solutions, which are incredibly expensive to purchase and maintain.One of the biggest drawbacks to Linux adoption in the commercial Enterprise space is its lack of SMP scalability. If the SGI platform works out we will start seeing Linux scaling into an arena that will allow for acceptance in the Enterprise.
In my universe I'm perfectly normal, it's not my fault you don't live in my universe.
That said, it's an impressive result. And it's done in an unusual way. SGI has a 1.6GB/s channel running through routers connecting the processors and memory. A computer is made up of multiple rackmount "bricks" connected by cables and routers. The "router" is a 2U rackmount device.
Processors and memory reside in rackmount boxes with 4 CPUs and 8 GB (max) of local memory. These boxes interconnect through a single 1.6GB/s link per box, which, in a big system, goes through several layers of routers. So a memory access to another box is routed through what is essentially a fast LAN. All this is cached, of course.
It's not clear to what extent application programs have to be aware of this. Clearly, if you lay things out in memory badly, with lots of CPUs reading and writing the same memory from all over the memory net, the system will bottleneck. (Everybody reading the same stuff is OK; it's cached. But writes have to propagate back to the home location of the data.)
Since the whole monster crashes all at once, you don't want to build your web server farm this way. It's for applications that really need all that crunch power in one machine.
I think it is pretty interesting that the benchmark that they used measured memory throughput, as opposed to, say, an actual workload. In other words, this is a synthetic benchmark, versus a real-world benchmark. They say, "Look! We can do memory transfers really really fast!"
Unfortunately, memory transfers are not the world when it comes to large multiprocessor boxes. The overhead comes in when you're trying to synchronize a large number of threads/CPUs to do a large task. For example, an Oracle database.
Sun has proven that it scales up the tree very well with large numbers of processors. But from my understanding, Linux is more efficient with a low processor count, and less and less efficient with more processors.
I question its ability to do anything with a real workload. And I've even more suspicious because they use a benchmark I've never heard of (STREAM TRIAD) to push its superiority on a single-aspect synthetic benchmark.
Good. The machine looks like it has a decent memory bus, and memory modules with a good configuration and speed rating. Now, what can the machine actually do well that makes it a real winner?
The whole point with the SGI supercomputers (there are Origin servers running Irix on 1024 processors) is that there's one single copy of the OS running across all those CPUs, and the entire memory is available to all CPUs on the same piece of hardware. That means, any CPU can access any piece of information at the speed of mem-IO, and you can easily create a large matrix (think many tens or hundreds of GB) to keep all your data in one piece.
Networked clusters (Mosix, Beowulf) split the CPU bunch across the network, and the memory is split too. That means there's a huge latency when a CPU wants to access data that happens to be on a different node on the network: the network latency is many times larger than memory latency.
There are problems that simply cannot be solved on networked clusters, precisely because of network latency. While true supercomputers (all CPUs on the same machine) do not have this limitation.
Well, ok, so you can split the matrix across nodes in a Beowulf, but even if you have the same CPU power as the SGI supercomp, you're going to solve the problem several times slower (if not several orders of magnitude slower). Such is the importance of latency.
This is why there's no point in clusterising this kind of computers: you lose their biggest advantage: single OS copy, all memory on the same machine.